Seedream / Dola Seedream 图像生成/编辑

Seedream / Dola Seedream 图像生成/编辑 is the pack's text-to-image / image-edit workhorse, and its neat trick is that it actually bundles two different closed models behind one dropdown. ByteDance's Seedream v5 Pro is a closed image model you cannot run locally - it sits on the official ComfyUI Partner Nodes roster too, so this isn't an obscure endpoint, it's the real deal accessed through this pack's reseller gateway. The model_family selector flips between the domestic seedream-v5-pro and the overseas dola-seedream-5.0-pro routes, each of which branches into a text-to-image and an image-edit variant depending on what you connect.

The mode-switching is the part to internalize: no reference image connected = text-to-image; any reference image connected = image editing. You don't flip a mode toggle, you just connect or disconnect image1. The node submits the matching model (...-t2i vs ...-i2i) automatically.

Inputs that matter:

• prompt - required, 5 to 2,000 characters.
• model_family - seedream-v5-pro (domestic) or dola-seedream-5.0-pro (overseas). Defaults to domestic.
• resolution - 1k, 2k, or custom. Pick custom and the width/height fields (240–8192) take over; the presets ignore them.
• output_format - png or jpeg. PNG if you might edit further, JPEG if you're shipping to a chat box.
• image1 … image10 - optional editing references. Connect any of them and you've switched to edit mode; the model follows all ten in slot order.

Outputs are the pack's standard image quartet: image (the real IMAGE tensor), image_url, task_id, and response.

How it works

Same upload → submit → poll → download pipeline as every node in this pack. Your reference images (if any) get uploaded to the api.seedance.nz gateway, the request is built with the right t2i/i2i model name, the remote job runs, and the finished picture streams down and decodes into a ComfyUI IMAGE. There's a seed control, but note the tooltip: it's a ComfyUI cache key - models without documented seed support don't receive it. That's the pack's universal seed behavior, not a Seedream quirk.

Installing it

The whole pack installs together: ComfyUI Manager → search ComfyUI Seedance, or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/ComfyUI_Seedance.git
python -m pip install -r ComfyUI_Seedance/requirements.txt

Restart, look under Seedance. There's nothing to download for this one - the dependency list is a single requests line and the model runs on the vendor's machines. You supply an API key from api.seedance.nz, either through the Seedance API Config node or the SEEDANCE_API_KEY env var.

Where people get burned

• Forgetting which mode you're in. You connect a reference image and expect a clean text-to-image render; instead the model obediently edits the reference. Or the reverse - you meant to edit but left all ten slots unconnected and got a fresh generation. The behavior is predictable; you just have to remember the rule.
• Overseas vs domestic isn't cosmetic. The routes go through different infrastructure, which can mean different pricing, latency, and moderation posture. If a prompt gets refused on domestic, the overseas route is the pack's documented escape hatch - the README's example set even includes "宽审核" (looser-audit) workflows.
• The api_key box vs the prompt box. Paste the prompt into prompt. The key field validates and rejects non-ASCII, but the error is friendlier if you just don't do it.

Standard fine print for the pack applies: results links expire, so save the image output, and your prompt plus reference images go to a third-party endpoint. Cost is per call, so do your resolution experiments at 1k.